Researchers at DoubleAI reported a 'mechanism intervention' method applied to the 763-billion-parameter DeepSeek-V4.1-Flash: the model reads its own internal activations from the deepest seven layers while solving math problems, then autonomously writes a probe and resampling routine to roll back and recompute when its reasoning goes off track.
On AIME and HMMT 2026, accuracy rose from 63.9% to 71.0% - over 7 percentage points - without modifying any weights. Qwen3-8B showed a similar roughly 7-point gain under zero-gradient updates, suggesting the approach generalizes across models.
